Architecture for Guaranteed Delay Service in High Speed Networks

Thumbnail Image


PhD_97-3.pdf (702.83 KB)
No. of downloads: 2223

Publication or External Link






The increasing importance of network connections coupled with the lack of abundant link capacity suggests that the day when service guarantees are required by individual connections is not far off. In this dissertation we describe a networking architecture that can efficiently provide end-to-end delay guarantees on a per- connection basis. In order to provide any kind of service guarantee it is imperative for the source traffic to be accurately characterized at the ingress to the network. Furthermore, this characterization should be enforceable through the use of a traffic shaper (or similar device). We go one step further and assume an extensive use of traffic shapers at each of the network elements. Reshaping makes the traffic at each node more predictable and therefore simplifies the task of providing efficient delay guarantees to individual connections. The use of per-connection reshapers to regulate traffic at each hop in the network is referred to as a Rate Controlled Service (RCS) discipline. By exploiting some properties of traffic shapers we demonstrate how the per-hop reshaping does not increase the bound on the end-to-end delay experienced by a connection. In particular, we show that an appropriate choice of traffic shaper parameters enables the RCS discipline to provide better end-to- end delay guarantees than any other service discipline known today. The RCS discipline can provide efficient end-to-end delay guarantees to a connection; however, by definition it is not work-conserving. This fact may increase the average delay that is observed by a connection even if there is no congestion in the network. We outline a mechanism by which an RCS discipline can be modified to be work-conserving without sacrificing the efficient end-to-end delay guarantees that can be provided to individual connections. Using the notion of service curves to bound the service process at each network element, we are able to provide an upper bound on the buffers required to ensure zero loss at the network element. Finally, we examine how the RCS discipline can be used in the context of the Guaranteed Services specification that is currently in the process of being standardized by the Internet Engineering Task Force.